Why design systems usually wait
Design systems have traditionally come late, and for good reason. Before you know what the product is, half the components you define get thrown away, and a system that's too complete becomes a cage: every new idea has to bend to structures you committed to before you understood the problem. The team was three full-stack engineers who leaned backend, and me as the only product designer. By the old math, the advice was obvious: ship the rough version first and systematize later.
What AI changed
Coding agents changed that math in both directions. Building got cheap: with a defined component to implement against, a polished screen ships about as fast as an improvised one. But the risk moved. Without a system to point at, an agent will cheerfully generate a slightly different button forty times, every screen locally reasonable while the product as a whole drifts apart. Under-investing in the system stopped saving time; it just handed the design decisions to the model.
Striking the balance
So the system came earlier than it traditionally would, but not all at once. I fully specified what the agents needed to stay coherent: the foundations, the components every screen shares, and the task shell, down to their states and naming, because those are exactly the decisions a model can't make for you. Anything still in question stayed deliberately thin, and product-specific components were added as the product earned them rather than designed up front in anticipation.
The test for each piece was simple: would leaving it undefined let the AI drift, or would defining it now lock us into a guess? The first got built. The second waited.
From Figma to code
We built that core in Figma, using Figma MCP to connect Claude Code directly into the library, with Code Connect providing the mapping layer, so a component defined once is what the engineers actually get in code. The team spent far less time on frontend UI and could focus on backend and business logic, which is what they wanted to be doing anyway.
Human-designed, AI-developed
The path from there was conventional even if the tooling wasn't. Wireframes first, to iterate early with the clinical and operations teams on what they actually needed. Then pixel-perfect mockups built on the system, with interactive prototypes covering real use cases end-to-end for user testing.
Polish isn't vanity in healthcare. There's an enormous amount of information in a clinical model, it's easy to hit information overload, and a poorly designed platform is exhausting to work in even when the workflows underneath it are right.